On May 14, 2026, a company not yet widely known but poised to disrupt the entire AI chip industry—Cerebras Systems—completed the largest tech IPO of the year on Nasdaq. Priced at $185 per share, the stock opened with a gap up to $350 and closed its first day up 68%. Dubbed "NVIDIA’s strongest challenger," this rising star in AI chips has launched a direct technological assault on GPU giant NVIDIA with its "dinner plate-sized" wafer-scale chip.
Just over a month after going public, however, Cerebras’ first quarterly report sparked sharp market divisions. Revenue exceeded expectations and losses narrowed significantly, but guidance for a steep drop in gross margin sent the after-hours stock price down more than 10%. What is the market worried about? Does the independent wafer-scale chip strategy truly have the long-term potential to challenge NVIDIA? We’ll systematically analyze Cerebras from four perspectives: technology, financial performance, industry competition, and trading channels.
Cerebras’ Value Proposition
Traditional chip manufacturing follows a familiar logic: a 12-inch silicon wafer is patterned with hundreds of chips, then cut, packaged, and tested. The size of each chip is limited by the photomask dimensions of the lithography machine, so chips can’t be made larger. Cerebras has upended this paradigm—rather than slicing up the wafer, it completes full-wafer lithography in one go, transforming an entire silicon wafer into a single, massive chip.
This is the Cerebras Wafer-Scale Engine (WSE). The latest generation, WSE-3, is built on TSMC’s 5nm process, with a single chip area of 46,225mm², integrating 4 trillion transistors and 900,000 AI cores, equipped with 44GB of on-chip SRAM, and delivering 125 petaflops of AI compute. For comparison, NVIDIA’s H100—the workhorse of AI data centers—contains about 80 billion transistors. WSE-3 packs 50 times more.
But the sheer transistor count isn’t Cerebras’ true moat. The real differentiation lies in its memory architecture.
Traditional GPUs (like the H100) rely heavily on external HBM high-bandwidth memory, with data transfers between chip and memory constrained by physical bandwidth. This is known in the industry as the "memory wall"—no matter how powerful the compute units are, if data can’t move fast enough, performance suffers. Cerebras integrates 44GB of SRAM directly onto the chip, achieving an on-chip memory bandwidth of 21 PB/s. Analysts note that WSE-3’s memory bandwidth is 2,625 times that of NVIDIA’s B200. In AI inference scenarios, this means model weights don’t need to be constantly shuttled from off-chip memory, drastically reducing inference latency.
Of course, the wafer-scale approach comes with trade-offs. A single fatal defect anywhere on the wafer can compromise the entire chip’s usability. Cerebras’ solution is "redundant core repair technology"—designing a large number of spare compute cores that automatically bypass defective regions. This increases design complexity and cost. Using the entire wafer as a single chip requires fundamentally different defect tolerance and yield management compared to traditional chip manufacturing.
Fundamental difference in technical approach: NVIDIA pursues "large-scale clustering + high-speed interconnects," assembling supercomputers from countless GPUs. Cerebras takes the "single-chip extreme scaling" route, replacing hundreds or thousands of GPUs with one giant chip. NVIDIA has decades of ecosystem and unmatched software compatibility. Cerebras offers theoretical efficiency advantages in specific inference scenarios, but its software ecosystem must be built from scratch.
First Month After IPO: From Euphoria to Correction—CBRS Price Trajectory
On May 14, 2026, Cerebras debuted on Nasdaq at $185 per share, opening at $350. The stock surged over 108% intraday, triggering a trading halt, and closed at $311.07. With a 68% first-day gain, it became the largest tech IPO in the US for 2026.
In the weeks that followed, CBRS shares saw wild swings. The price reached a historic high of $386, then dropped to a low near $197. As of Tuesday, June 23, CBRS closed at $226.72—still about 23% above the IPO price, but down more than 27% from its first-day close.
After the June 23 close, Cerebras released its first quarterly earnings as a public company, sending the after-hours price tumbling over 10%. On June 24 night trading, CBRS fell nearly 11% further, to $201.8.
As of publication, CBRS’ market cap is about $49.8 billion, with a trailing P/E ratio of roughly 527x. This valuation reflects high growth expectations, but also means any disappointment can trigger sharp volatility.
First Earnings Report: Revenue Beats, But Why Is the Market Unimpressed?
Cerebras’ Q1 2026 (ending March 31) financials showed a striking "two-sided" picture:
Positive surprises:
- Total revenue of $193.4 million, up 94% year-over-year, beating analyst expectations of $181.2 million
- Hardware revenue of $110.6 million, up 59% year-over-year
- Cloud and other services revenue of $82.8 million, up 178% year-over-year
- Net loss of $14 million, sharply narrowed from $23.9 million a year ago
- Loss per share of $0.22, better than the expected $0.25
- Full-year revenue guidance of $855–$865 million, above the analyst consensus of $824.8 million
Concerns:
- Q2 core gross margin guidance at 36%-38%, down more than 10 percentage points from Q1’s 46.5%
- Full-year core operating margin expected at -28% to -32%
- Q2 core revenue expected at around $194 million
Revenue doubling, narrowing losses, and raised guidance—all impressive for any growth company. Yet the market’s response was a steep after-hours selloff. The logic is straightforward: Cerebras’ valuation is built on a dual expectation of "high growth + high margins," and the sharp drop in gross margin undermines the latter.
CFO Bob Komin explained on the earnings call that the margin decline was due to a shortage of data center space, forcing Cerebras to lease back some systems from customers, while the company is "actively" expanding its own capacity. These costs are expected to reduce 2026 margins by about 10–15 percentage points. CEO Andrew Feldman was blunt: "It’s deeply ironic that after inventing all this technology, the limiting factor is actually building the buildings."
In other words, Cerebras’ current bottleneck isn’t technology or demand, but the pace of physical infrastructure supply lagging order growth. This creates short-term profit pressure, but also confirms the reality and urgency of demand.
OpenAI and AWS: Customer Base Transformation Behind $20 Billion Orders
The evolution of Cerebras’ customer base is key to understanding its long-term value.
In the first half of 2024, UAE-based AI company G42 accounted for 87% of Cerebras’ revenue. Such extreme customer concentration was once the market’s top concern. But in January 2026, Cerebras announced a strategic partnership with OpenAI worth over $20 billion—OpenAI will deploy 750 megawatts of Cerebras high-speed inference compute by 2028. The two also jointly launched Codex-Spark, an AI model designed for near-instant coding, generating over 1,000 tokens per second.
Meanwhile, Cerebras established a multi-year strategic partnership with Amazon AWS, planning to deploy Cerebras CS-3 systems in AWS data centers. The two will offer a "decoupled inference strategy": AWS’s Trainium 3 chips handle prefill compute, while Cerebras CS-3 delivers ultra-fast decoding inference.
These partnerships mean far more than just order size. From a single customer (G42) to dual pillars (OpenAI and AWS), Cerebras’ customer concentration risk has been substantially mitigated. More importantly, OpenAI and AWS represent the two core global AI inference scenarios—cutting-edge model training and massive-scale cloud deployment. Securing long-term orders from both giants is itself a "market validation" of Cerebras’ technology path.
By the end of 2025, Cerebras had $24.6 billion in backlog contracts, with the company expecting to convert $3.7 billion of that to recognized revenue by 2027. The ratio of backlog to current revenue is about 48x—this shows future revenue visibility, but also that Cerebras is still in the early stages of large-scale delivery.
Wafer-Scale Chips: The Rationale and Limits of Challenging NVIDIA’s Monopoly
Cerebras has chosen a technical path fundamentally different from NVIDIA.
NVIDIA follows the industry mainstream—chiplet assembly. Chips are split into compute, cache, IO, and other functional "chiplets," manufactured separately and then joined via advanced packaging. This approach offers high yield, controllable costs, and scalable mass production. NVIDIA’s B200 and Huawei’s Ascend chips use this method.
Cerebras’ wafer-scale route is "cast as one"—no slicing, no assembly, the entire wafer is used as a single chip. This offers theoretical efficiency advantages for inference, but faces high manufacturing complexity, tough yield management, and the challenge of building a software ecosystem from scratch.
Their competition is essentially a battle between "scale effects" and "extreme efficiency" paradigms. NVIDIA’s strengths are its decades-old CUDA software ecosystem and massive production capacity. Cerebras’ advantage is the potential for 10x speed gains in specific inference scenarios.
For investors, the key question isn’t "Can Cerebras beat NVIDIA?"—that’s nearly impossible in the foreseeable future. The real question is: Is the AI inference market big enough to support an independent, non-GPU technical path? If yes, then Cerebras’ uniqueness as the only commercial player on this path becomes a valuation logic in itself.
Risk Factors: Four Major Challenges Not to Overlook
Uncertainty in gross margin and profitability. Q2 margin guidance dropped from 46.5% to 36%-38%, with full-year core operating margin still deeply negative. The company is far from sustainable profitability. Morgan Stanley believes margin compression is temporary, expecting margins to return to the 60% target as Cerebras moves away from leased infrastructure—but this remains unproven in the market.
Structural changes in customer concentration need time to validate. While OpenAI and AWS have greatly improved customer diversification, OpenAI’s $20 billion order still dominates the backlog ($20 billion out of $24.6 billion). Any change in OpenAI’s deployment pace could significantly impact revenue.
Supply pressure from lockup expiration. This Thursday (June 25), the lockup period expires, freeing about 13% of IPO shares for sale by early backers and insiders. Increased float could create short-term price pressure.
Valuation-growth mismatch. CBRS currently trades at about 91x sales, far above NVIDIA’s 23x. High-growth companies deserve a premium, but if growth slows or margins fail to improve, the risk of valuation contraction is significant.
Conclusion
Cerebras’ rise is a microcosm of AI compute demand shifting from "training" to "inference." As large model training becomes standardized and scaled, the drive for extreme latency, cost, and energy efficiency in inference is opening a commercial window for "non-mainstream" technologies like wafer-scale chips.
From its first earnings report, Cerebras delivered revenue and orders that beat expectations, but the sharp drop in margins also exposed the growing pains of early-stage expansion—physical data center infrastructure can’t keep up with the explosion in compute demand. It’s a "sweet headache," but also a real drag on profits.
For investors, Cerebras isn’t "NVIDIA’s replacement," but "another possibility for AI inference." The ultimate outcome of this independent path depends on two core variables: whether the AI inference market continues to expand rapidly enough to support multiple technical routes, and whether Cerebras can efficiently convert its $24.6 billion backlog into revenue and positive cash flow in 2026–2027.
Can Cerebras’ wafer-scale chip truly shake NVIDIA’s GPU empire? The answer may not be clear today, but will unfold at each critical milestone over the next 12–24 months: order conversion rates, margin recovery, and AWS partnership implementation.
FAQ
Q1: What is the core difference between Cerebras’ wafer-scale chip WSE-3 and NVIDIA’s H100?
WSE-3 is a single, uncut 12-inch wafer integrating 4 trillion transistors and 900,000 cores; H100 is a traditionally cut and packaged chip. The main difference is memory architecture: WSE-3 has 44GB of on-chip SRAM with 21PB/s bandwidth; H100 relies on external HBM with only 3.35TB/s bandwidth. WSE-3 offers significant speed advantages for inference, but comes with higher manufacturing complexity and cost.
Q2: What are the key numbers from Cerebras’ Q1 2026 earnings report?
Q1 revenue was $193.4 million, up 94% year-over-year and beating the expected $181.2 million; net loss was $14 million, sharply narrowed from $23.9 million a year ago; hardware revenue was $110.6 million, cloud services revenue was $82.8 million. Full-year revenue guidance is $855–$865 million.
Q3: Why did Cerebras’ stock drop after the earnings report?
Despite revenue and loss beating expectations, Q2 gross margin guidance dropped sharply from Q1’s 46.5% to 36%-38%. The main reason is a shortage of data center space, forcing the company to lease back systems from customers and aggressively expand capacity, with related costs reducing margins by about 10–15 percentage points. The market is concerned about the visibility of the path to profitability.
Q4: What are Cerebras’ main risks?
Four major risks: sharp drop in gross margin and uncertain profitability; customers still highly concentrated in OpenAI ($20 billion order out of $24.6 billion backlog); lockup expiration on June 25, freeing about 13% of IPO shares for sale; price-to-sales ratio around 91x, far above NVIDIA’s 23x, creating significant valuation contraction risk.
